On local Turán density problems of hypergraphs
Chunqiu Fang
Guorong Gao
Jie Ma
Ge Song
School of Computer Science and Technology, Dongguan University of Technology, Dongguan, Guangdong 523808, China. Email: [email protected] of Mathematical Sciences, University of Science and Technology of China,
Hefei, Anhui 230026, China. Email: [email protected] of Mathematical Sciences, University of Science and Technology of China,
Hefei, Anhui 230026, China. Email: [email protected] of Management, University of Science and Technology of China,
Hefei, Anhui 230026, China. Email: [email protected].
Abstract
For integers q≥p≥r≥2, we say that an r-uniform hypergraph H has property (q,p), if for any q-vertex subset Q of V(H), there exists a p-vertex subset P of Q spanning a clique in H. Let Tr(n,q,p)=min{e(H):H⊂(r[n]),H has property (q,p)}. The local Turán density about property (q,p) in r-uniform hypergraphs is defined as tr(q,p)=limn→∞Tr(n,q,p)/(rn). Frankl, Huang and Rödl [J. Comb. Theory, Ser. A, 177 (2021)] showed that limp→∞tr(ap+1,p+1)=ar−11 for positive integer a and t3(2p+1,p+1)=41 for all p≥3 and asked the question that determining the value of limp→∞tr(γp+1,p+1), where γ≥1 is a real number. Based on the study of hypergraph Turán densities, we determine some exact values of local Turán densities and answer their question partially; in particular, our results imply that the equality in their question about exact values does not hold in general.
1 Introduction
A hypergraph H is a pair (V,E), where V is a finite set of elements, called vertices, and E is a set of nonempty subsets of V, called hyperedges. The hypergraph H is said to be r-uniform if every hyperedge of H is of size r. The simple graphs are the 2-uniform hypergraphs. We denote the numbers of the vertices and the hyperedges of H by ν(H) and e(H), respectively. Let (rV(H)) be the family of all r-element subsets of V(H). For an r-uniform hypergraph H, we use H to denote the complement of H, that is V(H)=V(H) and E(H)=(rV(H))∖E(H). For a vertex set U⊂V(H), denote by H[U] the subhypergraph induced by U, that is H[U]={e∈E(H):e⊂U}. A set I of vertices in a hypergraph H is independent if H[I] does not contain a hyperedge. The independence number of H, denoted by α(H), is the maximum cardinality of an independent set in H. A vertex set X in a hypergraph H is a clique if H[X] is a complete r-uniform hypergraph. The clique number of H is the maximum cardinality of a clique in H. For two vertices x,y∈V(H), we denote dH(x)=∣{e∈E(H):e∋x}∣ and dH(xy)=∣{e∈E(H):e⊃{x,y}}∣. Throughout this paper, we write [n]={1,2,…,n}.
1.1 Turán problems
The study of Turán number is one of the central topics in extremal graph (hypergraph) theory. Let F be a family of r-uniform hypergraphs, we say that an r-uniform hypergraph H is F-free if H contains no member of F as a subhypergraph. The Turán number of F, denoted ex(n,F), is the maximum number of hyperedges in an F-free r-uniform hypergraph on n vertices. We let π(F):=limn→∞(rn)ex(n,F) and call π(F) the Turán density of F.
Katona, Nemetz and Simonovits [14] used an averaging argument to show that the Turán density of any family of hypergraphs exists. If F={F}, we denote the Turán number ex(n,{F}) by ex(n,F) and Turán density π({F}) by π(F), respectively.
The classic Turán Theorem [21] says that for p≥2, the Turán number ex(n,Kp+1) is uniquely attained by the complete balanced p-partite graph on n vertices. The following celebrated Erdős-Stone-Simonovits Theorem [6, 7] gives a tight estimate for the Turán number of any family of graphs.
Theorem 1.1** (Erdős-Stone-Simonovits, [6, 7]).**
For any family F of non-empty graphs, we have
[TABLE]
where p=Ψ(F)=min{χ(F):F∈F}−1 denotes the subchromatic number of F.
Note that Theorem 1.1 implies that π(F)=χ(F)−1χ(F)−2 for any non-empty graph F. Extending Turán’s Theorem to r-uniform hypergraphs is one of the most challenging problems in extremal graph theory. Erdős offered a money prize for determining π(Kkr) for at least one pair r,k with k>r≥3. However, no Turán density π(Kkr) is known for any k>r≥3 yet.
Conjecture 1.2** (Turán, [21]).**
For every integer k≥4,
[TABLE]
The most famous case k=4 in Conjecture 1.2, which declares that π(K43)=95 has attracted a lot of interest and activity through the years. It is known that many constructions (e.g. see [3, 8, 15]) achieve the conjectured value. To date, Razborov [19] proved the best known upper bound π(K43)≤0.561666 by using flag algebra.
1.2 Local Turán density
We consider local Turán densities which are defined as follows.
Definition 1.3** (Frankl-Huang-Rödl, [11]).**
For integers q≥p≥r≥2, we say that an r-uniform hypergraph H=(V,E) has property (q,p), if for every vertex set Q⊂(qV), there exists a vertex set P⊂(pQ) spanning a clique in H, that is, (rQ)⊂H.
Let Tr(n,q,p)=min{e(H):H⊂(r[n]),Hhas property(q,p)}.
Let tr(n,q,p)=(rn)Tr(n,q,p) and tr(q,p):=limn→∞tr(n,q,p).
We call tr(q,p) the local Turán density of property (q,p) in r-uniform hypergraphs.
For q≥p≥r≥2, let Gq,pr be the family of r-uniform hypergraphs G with ν(G)=q and α(G)≤p−1.
Note that an r-uniform hypergraph H has property (q,p) if and only if H is Gq,pr-free. Then we have Tr(n,q,p)+ex(n,Gq,pr)=(rn). Thus,
[TABLE]
This also shows that the local Turán density exists. By the definition of local Turán density, for any integers q>p>r≥2,
tr(q,p+1)≥tr(q+1,p+1)≥tr(q,p)≥tr(q+1,p).
In [5], it was shown that for r=2 (the graph case), t2(q,p)=1/⌊p−1q−1⌋. For general r, Frankl and Stechkin [13] showed that if q≤r−1r(p−1), then tr(q,p)=1.
Frankl [10] proved that limp→∞t3(2p+1,p+1)=41.
Frankl, Huang and Rödl [11] generalize it to the r-uniform hypergraphs. They proved the following Theorem.
Theorem 1.4** (Frankl-Huang-Rödl, [11]).**
For integers r≥2 and a≥2,
[TABLE]
In the same paper, they [11] obtained the exact value of t3(2p+1,p+1) for all p≥3.
Theorem 1.5** (Frankl-Huang-Rödl, [11]).**
For every integer p≥3,
[TABLE]
In light of these results, they ask the the following question.
Question 1.6** (Frankl-Huang-Rödl, [11]).**
Is it possibly true that for every positive real number γ>1,
[TABLE]
where F is the family of all the r-uniform hypergraphs satifying ν(F)≥γα(F)?
1.3 Our results
In this section we introduce our main results. Our first result Theorem 1.7 gives a weaker version of the first equality of Question 1.6,
while the second result Theorem 1.8 is about some exact values of local Turán densities, which implies that the second equality in Question 1.6 for limit values does not hold in general (on the other hand, it does hold for several sub-intervals).
Theorem 1.7**.**
Let r≥3 be an integer and γ>1 be a real number. We have
[TABLE]
where Fγr={r-uniform graphs F:ν(F)>γα(F)}. Furthermore, for any integer r≥3, there exist a real number γ and a positive integer p0=p0(r,γ), such that for any integer p≥p0, we have tr(⌊γp⌋+1,p+1)=1−π(Fγr).
Remark 1.
From Theorem 1.7, after replacing the hypergraph family F in Question 1.6 by Fγr, if we can show π(Fγr)=minF∈Fγrπ(F),111For a general hypergraph family G, it is known that π(G)=minG∈Gπ(G) does not hold. then the first equality of Question 1.6 holds.
Theorem 1.8**.**
The following results about local Turán density hold.
(1) For a fixed positive real number γ, there exists a positive integer p0=p0(γ), such that for all integers p≥p0, we have the following
[TABLE]
(2) Let γ be a positive real number with 34≤γ<57, there exists a positive integer p0=p0(γ) such that for all integers p≥p0, we have t4(⌊γp⌋+1,p+1)=1−444!.
(3) For integers r≥3 and p≥r2−r−1, tr(rp+1,(r−1)p+1)=1−rrr!.
Remark 2.
We point out that strictly speaking the first equality in Question 1.6 may not be true. For example, let r=3 and γ=23. Then the hypergraph family F in Question 1.6 contains the single edge as its member. Thus 1−minF∈Fπ(F)=1. But Theorem 1.8 shows that
t3(⌊23p⌋+1,p+1)=97, which implies that the first equality of Question 1.6 fails in this case.
In Theorem 1.7, we modify the corresponding definition to be ν(F)>γα(F).
Remark 3.
It can be seen from the proofs that the conclusions of Theorems 1.7 and 1.8 still hold when replacing ⌊γp⌋ with ⌈γp⌉.
The rest of the paper is organized as follows. In Section 2, we prove Theoem 1.7. In Section 3, we give some extremal constructions for local Turán densities and prove Theorem 1.8. In Section 4, we conclude with some remarks and nature problems on this topic.
2 Proof of Theorem 1.7
In this section, we will prove Theorem 1.7. The method comes from the proof of Theorem 1.4 in [11]. For the integer pair (q,p) with q≤γp, we let e(q,p)=γp−q. Note that since q≥p, we always have e(q,p)≤γq−q=(γ−1)q. We also need the following definition and lemmas.
Definition 2.1** (Frankl-Huang-Rödl, [11]).**
For H⊂(rX), the vertex set Z⊂X is a (w,v)-hole if ∣Z∣=w>γv and the clique number of H[Z] is v.
Lemma 2.2** (Frankl-Huang-Rödl, [11]).**
Suppose H⊂(rX) has property (q,p) and Z is a (w,v)-hole of H with w<q, then H[X∖Z] has property (q−w,p−v).
Proof.
Take an arbitrary set U∈(q−wX∖Z), then U∪Z∈(qX). Since H has property (q,p), H[U∪Z] contains a clique of size p. Hence H[U] contains a clique of size p−v. Then H[X∖Z] has property (q−w,p−v).
∎
Lemma 2.3**.**
For every r≥3, γ>1 and ϵ>0, there exists ℓ0=ℓ0(r,ϵ,γ) such that the following holds for all ℓ≥ℓ0. Suppose an r-graph H on vertex set X has property (q,p) for all pairs (q,p) with q≤ℓ, p=⌈γq⌉ (In other words, H does not have a (w,v)-hole with ℓ≥w>γv). Then for all Y∈(lX),
[TABLE]
Proof.
Since
limℓ→∞(rℓ)ex(ℓ,Fγr)=π(Fγr),
there exists ℓ0=ℓ0(r,ϵ,γ) such that for all ℓ≥ℓ0,
ex(ℓ,Fγr)≤(π(Fγr)+ϵ(1−π(Fγr)))(rℓ).
For Y⊂(ℓX), H[Y] has property (q,p) for all pairs (q,p) with q≤ℓ, p=⌈γq⌉, which means that H[Y] is Fγr-free. Thus, we have
[TABLE]
∎
Now we are ready to prove Theorem 1.7.
Proof of Theorem 1.7.
Let H be a maximum Fγr-free r-graph on n vertices, then H has property (⌊γp⌋+1,p+1) (also each of the properties (⌈γp⌉+1,p+1), (⌊γp⌋,p), and (⌈γp⌉,p)). Thus, tr(⌊γp⌋+1,p+1)≤1−π(Fγr).
Now we focus on the lower bound. Given ϵ>0, let us fix a large integer ℓ≥ℓ0=ℓ0(r,2ϵ,γ), where ℓ0(r,2ϵ,γ) is obtained from Lemma 2.3. Let
[TABLE]
And then fix a much larger integer L≥θ04γℓ2. Consider a sufficiently large r-uniform hypergraph H⊂(r[n]) having property (q,p), where q=⌊γL⌋ and p=L.
Our aim is to find a subset X⊂[n] with (r∣X∣)>(1−2ϵ)(rn) such that H[X] has no (w,v)-hole with w≤ℓ and v≥r−1.
We start with H0=H and define Hi inductively. Let q0=q, p0=p and X0=[n]. Suppose that Hi⊂(rXi) has property (qi,pi) and it still has a (wi,vi)-hole. Then we take such a (wi,vi)-hole Zi⊂Xi and set
[TABLE]
By Lemma 2.2, Hi+1 has property (qi−wi,pi−vi). Moreover,
[TABLE]
Set qi+1=qi−wi, pi+1=pi−vi and repeat. At every step, we have
[TABLE]
Since vi≥r−1 for all i and p0=p, we have i≤r−1p. Suppose at step i, the hypergraph Hi no longer contains a (w,v)-hole with w≤ℓ. Then we choose a subset Q of size ℓ of V(Hi) uniformly at random. By Lemma 2.3, we have
[TABLE]
On the other hand, ∣Xi∣≥n−iℓ≥n−r−1pℓ. Thus, for sufficiently large n, (r∣Xi∣)≥(1−2ϵ)(rn). Therefore,
[TABLE]
Otherwise suppose this process continues to produce (w,v)-holes. Let m be the first index such that qm<2l. Since e(qm,pm)≤(γ−1)qm and e(qi,pi) strictly increases at least θ0 after each step, it follows that m≤θ0(γ−1)qm. Thus,
[TABLE]
contradicting that L≥θ04γℓ2. We finish the proof of the first part of Theorem 1.7.
The proof of the other part is given by Theorem 1.8.
∎
3 Exact values of some local Turán densities
In this section, we determine some exact values of local Turán densities based on the known hypergraph Turán densities.
3.1 Extremal constructions
In this section, we construct two hypergraph families, which provide upper bounds for local Turán densities.
Definition 3.1**.**
For integers n,a≥1,r≥3 and k≥1, let Kn,a,kr be the family of r-uniform hypergraphs H with ν(H)=n and V(H) admits a vertex partition
V(H)=V1∪V2∪⋯∪Va and Va=U0∪U1∪⋯∪Uk, such that
E(H)=(⋃i=1a−1(rVi))⋃(rVa∖U0)⋃(⋃j=1k(rU0∪Uj)).
Let ρr(n,a,k)=min{e(H):H∈Kn,a,kr} and ρr(a,k)=limn→∞(rn)ρr(n,a,k).
When k=1, it is the extremal construction considered in [11] and we have ρr(a,1)=ar−11.
We compute its exact value for the case r=3,a=1 by the following proposition.
Proposition 3.2**.**
Let k be an integer. We have ρ3(1,k)=9k5k+4.
Proof.
Let H∈Kn,1,k3. Let V(H)=U0∪U1∪⋯∪Uk and E(H)=(3V(H))∖⋃1≤i1<i2≤kU0×Ui1×Ui2.
We may assume that ∣U1∣=∣U2∣=⋯=∣Uk∣=xn. Then ∣U0∣=(1−kx)n and 0≤x≤k1. Thus e(H)=(3n)−(2k)(xn)2(1−kx)n. Denote f(x)=x2(1−kx) where 0≤x≤k1. By direct calculation, we have
fmax=f(3k2)=27k24.
Thus we have ρ3(1,k)=limn→∞(3n)(3n)−(2k)fmaxn3=9k5k+4.
∎
Theorem 3.3**.**
Let a,k≥1, r≥3 and p≥r−1 be integers. For large n and any r-uniform hypergraph H∈Kn,a,kr, H has property (⌊(a+1−k1)p⌋+1,p+1). Furthermore, we have
[TABLE]
Proof.
Let V(H)=V1∪V2∪⋯∪Va be the vertex set partition of V(H), where Va=U0∪U1∪⋯∪Uk. For any vertex set X∈(⌊(a+1−k1)p⌋+1V(H)), we will show that H[X] contains a clique of size p+1.
We may assume that ∣X∩Vi∣≤p for any 1≤i≤a−1, otherwise H[X∩Vi] contains a clique of size at least p+1. Thus, ∣X∩Va∣≥2p+1−⌊kp⌋. If p<2k, then ∣X∩Va∣≥2p, one can easily find a clique of size at least p+1 in H[X∩Va]. Now we assume that p≥2k. Simiarly, we may assume that p+1−⌊kp⌋≤∣X∩U0∣≤p, otherwise either H[X∩U0] or H[X∩(U1∪⋯∪Uk)] contains a clique of size at least p+1. Assume now that ∣X∩U0∣=p−⌊kp⌋+t, where 1≤t≤⌊kp⌋. Then ∣X∩(U1∪⋯∪Uk)∣≥p+1−t. By the pigeonhole principle, there is some j with 1≤j≤k such that ∣X∩Uj∣≥⌈kp+1−t⌉. Thus, ∣X∩(U0∪Uj)∣≥p−⌊kp⌋+t+⌈kp+1−t⌉≥p+1. Therefore H[X∩(U0∪Uj)] contains a clique of size at least p+1.
∎
Definition 3.4**.**
For integers n,a≥1,r≥3 and 2≤k≤r−1, let Ln,a,kr be the family of r-uniform hypergraphs H with ν(H)=n and V(H) admits a vertex partition
V(H)=V1∪V2∪⋯∪Va and Va=U0∪U1∪⋯∪Uk, such that
E(H)=(⋃i=1a−1(rVi))⋃(⋃j=0k(rVa∖Uj)).
Let ηr(n,a,k)=min{e(H):H∈Ln,a,kr}
and ηr(a,k)=limn→∞(rn)ηr(n,a,k).
We compute its exact value for the case a=1 by the following proposition.
Proposition 3.5**.**
Let r≥3 and 2≤k≤r−1. We have ηr(1,k)=∑i=1k(−1)i+1(k+1−ik+1)(k+1k+1−i)r.
Proof.
Let H∈Ln,1,k3 with the minmum number of hyperedges. Let V(H)=U0∪U1∪⋯∪Uk and E(H)=⋃j=0k(rV(H)∖Uj).
We may assume that ∣U0∣=∣U1∣=∣U2∣=⋯=∣Uk∣=k+1n.
By the inclusion-exclusion principle, we have
[TABLE]
Then the edge density of H is
[TABLE]
∎
Remark 3.6**.**
Let r≥3. We have ηr(1,r−1)=1−rrr!.
Theorem 3.7**.**
Let a≥1,r≥3, p≥r−1 and 2≤k≤r−1 be integers. For large n and any r-uniform hypergraph H∈Ln,a,kr, H has property (⌊(a+k1)p⌋+1,p+1). Furthermore, we have
[TABLE]
Proof.
Let V(H)=V1∪V2∪⋯∪Va be the vertex set partition of V(H), where Va=U0∪U1∪⋯∪Uk. For any vertex set X∈(⌊(a+k1)p⌋+1V(H)), we will show that H[X] contains a clique of size p+1.
We may assume that ∣X∩Vi∣≤p for any 1≤i≤a−1, otherwise X∩Vi contains a clique of size at least p+1. Thus, ∣X∩Va∣≥⌊k(k+1)p⌋+1≥k(k+1)p+1. By the pigeonhole principle, there is some j with 0≤j≤k such that ∣X∩(Va∖Uj)∣≥⌈k+1k∣X∩Va∣⌉≥p+1. Thus H[X] contains a clique of size at least p+1.
∎
3.2 Proof of Theorem 1.8
In this subsection, we prove Theorem 1.8.
The key tool we use is the “blow-up” operation of the r-uniform hypergraphs. Let H be an r-uniform hypergraph on vertex set {v1,v2,…,vl} and μ=(μ1,μ2,…,μl) be an integer vector with each μi≥1.
Then the blow-up H(μ) is the r-uniform hypergraph formed by replacing the vertex vi of H with a disjoint class of μi vertices for each i and inserting a complete r-partite r-uniform hypergraph between any vertex classes corresponding to an edge in H. Given a family H={H1,H2,…,Hs} of r-uniform hypergraphs and a family T={μ1,μ2,…,μs} of positive integer vectors with each μi of dimension ∣V(Hi)∣, we define the T-blow-up of H to be H(T)={Hi(μi):1≤i≤s}. Brown and Simonovits [4] proved the following extremely useful result.
Theorem 3.8** (Brown-Simonovits, [4]).**
If H={H1,H2,…,Hs} is a family of r-uniform hypergraphs and T={μ1,μ2,…,μs} is a family of positive integer vectors with each μi of dimension ∣V(Hi)∣, then π(H(T))=π(H).
We also need the following results on hypergraph Turán densities.
Theorem 3.9** (Frankl-Füredi, [9]).**
Let F5={123,124,345}, we have π(F5)=92.
Theorem 3.10** (Baber-Talbot, [1]).**
Let H1={123,124,134,234},H2={123,124,125,345,346},H3={123,124,345,156,256}, H4={123,124,125,346,356,456}, then π({H1,H2,H3,H4})=278.
Let H2−={123,124,125,345}. Note that H2 is a subgraph of some blow-up of H2−, thus we have the following corollary.
Corollary 3.11**.**
π({H1,H2−,H3,H4})=278.**
Theorem 3.12** (Mubayi-Rödl, [17]).**
Let H7=(3[4])∪{(a,x,y):a∈[4],x,y∈{5,6,7},x=y}∖{(1,5,6)}, then π(H7)=43.
For r≥3, let the generalized triangle Tr be the r-uniform hypergraph on (2r−1) vertices with three edges {1,…,r},{1,…,r−1,r+1}and{r,r+1,…,2r−1}. Pikhurko [18] obtained the Turán density of T4.
Theorem 3.13** (Pikhurko, [18]).**
π(T4)=444!.**
For r≥3, the generalized fan denoted by Fr is the r-uniform hypergraph comprising r+1 edges e1,…,er,e such that ei∩ej={x} for all i=j, where x∈/e and ∣ei∩e∣=1 for all i. Mubayi and Pikhurko [16] obtained the Turán density of Fr.
Theorem 3.14** (Mubayi-Pikhurko, [16]).**
π(Fr)=rrr!.**
Now we are ready to prove Theorem 1.8.
Proof of Theorem 1.8.
(1) If 23≤γ<35, by Theorem 3.3 and Proposition 3.2, t3(⌊γp⌋+1,p+1)≤t3(⌊23p⌋+1,p+1)≤ρ3(1,2)=97.
We just need to prove the lower bound. Let μ=(5⌊γp⌋+1,5⌊γp⌋+1,5⌊γp⌋+1,5⌊γp⌋+1,5⌊γp⌋+1)222When 5⌊γp⌋+1 is not an integer, we always let it be its floor or ceil such that the sum of all the elements in μ equals ⌊γp⌋+1. We do the same if similar situations occur in the rest of the paper..
By Theorem 3.8 and Theorem 3.9, we have π(F5(μ))=π(F5)=92. Also, α(F5(μ))≤53(⌊γp⌋+1)+O(1)<p+1 for large p. By (1), for large p, we have t3(⌊γp⌋+1,p+1)≥1−π(F5(μ))=97.
If 35≤γ<47, by Theorem 3.3 and Proposition 3.2, t3(⌊γp⌋+1,p+1)≤t3(⌊35p⌋+1,p+1)≤ρ3(1,3)=2719. We just need to prove the lower bound. Let
[TABLE]
Then for large p, α(H1(μ1))≤2⌊γp⌋+1+O(1)<p+1, α(H2−(μ2))≤74(⌊γp⌋+1)+O(1)<p+1, α(H3(μ3))≤74(⌊γp⌋+1)+O(1)<p+1, α(H4(μ4))≤74(⌊γp⌋+1)+O(1)<p+1. By (1), Theorem 3.8 and Corollary 3.11, for large p, we have t3(⌊γp⌋+1,p+1)≥1−π({H1(μ1),H2−(μ2),H3(μ3),H4(μ4)})=1−278=2719.
If 2≤γ<37, by Theorem 3.3 and Proposition 3.2, t3(⌊γp⌋+1,p+1)≤t3(2p+1,p+1)≤ρ3(2,1)=41.
We just need to prove the lower bound. Let μ=(7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1). By Theorem 3.8 and Theorem 3.12, we have π(H7(μ))=π(H7)=43. Also, α(H7(μ))≤73(⌊γp⌋+1)+O(1)<p+1 for large p. By (1), for large p, we have t3(⌊γp⌋+1,p+1)≥1−π(H7(μ))=41.
For 3≤γ<722, the proof will be given in subsection 3.3 (See Theorem 3.20).
(2) For 34≤γ<57, by Theorem 3.7 and Remark 3.6, t4(⌊γp⌋+1,p+1)≤t4(⌊34p⌋+1,p+1)≤η4(1,3)=1−444!. Now we focus on proving the lower bound. Let μ=(7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1,7⌊γp⌋+1). By Theorem 3.8 and Theorem 3.13, we have π(T4(μ))=π(T4)=444!. Also, α(T4(μ))≤75(⌊γp⌋+1)+O(1)<p+1 for large p. By (1), for large p, we have t4(⌊γp⌋+1,p+1)≥1−π(T4(μ))=1−444!.
(3) By Theorem 3.7 and Remark 3.6, tr(rp+1,(r−1)p+1)≤ηr(1,r−1)=1−rrr!. Now we focus on proving the lower bound.
For r≥3 and p≥r2−r−1, let Fpr be the blow-up of Fr by replacing each vertex of e with p−r2+2r vertices, the common vertex x with (r−1)2 vertices and each vertex of V(Fr)∖(e∪{x}) with r−1 vertices. By Theorem 3.8 and Theorem 3.14, we have π(Fpr)=π(Fr)=rrr!. Note that ν(Fpr)=rp+1. Also, α(Fpr)=(r−1)p. By (1), we have tr(rp+1,(r−1)p+1)≥1−π(Fpr)+O(1)=1−rrr!.
∎
3.3 Proof of Theorem 1.8 when 3≤γ<722
To obtain more exact values of local Turán densities, we consider an extension of Theorem 3.12. We define the family Ha of 3-uniform hypergraphs.
Definition 3.15**.**
For an integer a≥2, let P and Q be two disjoint vertex sets with ∣P∣=2a and Q=a+1. Let Ha be the 3-uniform hypergraph on vertex set P∪Q with E(Ha)=(3P)∪{xyz:x∈P,y,z∈Q}. Let
[TABLE]
Using Turán’s Theorem [21], we have the following lemma about Turán number in multigraphs.
Lemma 3.16**.**
Let n≥a≥2 be integers and G be a multigraph on n vertices. If any a+1 vertices of G will induce at least one edge with multiplicities at least 2, then e(G)≥a2(2n)−n.
Proof.
By Turán’s Theorem [21], ex(n,Ka+1)≤(1−a1)(2n)+2n. Thus T2(n,a+1,2)=(2n)−ex(n,Ka+1)≥a1(2n)−2n. Hence e(G)≥2×T2(n,a+1,2)≥a2(2n)−n.
∎
Now we will compute the Turán density of family Ha under the condition that a warm version of Turán’s Conjecture (Conjecture 1.2) holds.
Theorem 3.17**.**
Let a≥2 be an integer. If π(K2a3)≤1−a21, then π(Ha)=1−a21.
Proof.
Let Gn,a=(3[n])∖((3V1)∪(3V2)∪⋯∪(3Va)) with V1⊔V2⊔⋯⊔Va=[n] and Vi∈{⌊an⌋,⌈an⌉}. It is easy to check that Gn,a is Ha-free and e(Gn,a)=(1−a21+o(1))(3n) as n→∞. Thus π(Ha)≥1−a21.
For the upper bounds, we just need to show that for any real number 0<ε<a21, there exist an integer N=N(ε) and a real number C=C(N), such that for any integer n≥N, ex(n,Ha)<(1−a21+ε)(3n)+an2+C. We prove it by induction on n≥N. Since π(K2a3)≤1−a21, we can choose a large integer N1=N1(ε) such that ex(n,K2a3)≤(1−a21+ε)(3n) holds for n≥N1. Furthermore, we can choose an integer N=N(ε)≥N1 and a real number C such that
(1−a21+ε)(3N)+aN2+C=(3N) and (1−a21+ε)(3n)+an2+C<(3n) holds for n>N.
It is easy to see that our claim holds for n=N. Now let n>N and assume that our claim holds for n−1. Let G be a 3-uniform hypergraph on vertex set [n] with e(G)≥(1−a21+ε)(3n)+an2+C. Therefore, G contains a clique with 2a vertices. By symmetry, suppose (3[2a])⊂G. For i∈[2a], we define the link graphs G(i)={{x,y}∈(2[2a+1,n]):ixy∈G}. Let M be the multigraph whose edge set is the union (with multiplicities) G(1)∪G(2)∪⋯∪G(2a). Let M=G(1)∪G(2)∪⋯∪G(2a). If there exists an (a+1)-subset S⊂[2a+1,n], such that the multiplicity of each edge of the induced subgraph of M on S is no more than 1, in other words, the multiplicity of each edge of the induced subgraph of M on S is at least 2a−1, then vertex set [1,2a]∪S induces some copy of hypergraph in Ha in G.
Thus, we assume that the induced subgraph of M on any (a+1)-subset of [2a+1,n] will contains at least one edge with multiplicities at least 2. By Lemma 3.16, e(M)≥a2(2n−2a)−(n−2a). Thus, e(M)≤(2a−a2)(2n−2a)+(n−2a). Therefore, there exists a vertex i∈[2a], such that e(G(i))≤2a1e(M)≤(1−a21)(2n−2a)+2an−1.
It follows that
[TABLE]
By induction hypothesis, we have
[TABLE]
So G−i contains a copy of some hypergraph in Ha. It follows that
[TABLE]
∎
Theorem 3.18**.**
Let a≥2 be an integer. If π(Ha)=1−a21, then for any real number γ with a≤γ<a+2a+11, there exist a postive integer p0=p(γ), such that for all integers p≥p0,
[TABLE]
Proof.
By Theorem 3.3, t3(⌊γp⌋+1,p+1)≤t3(ap+1,p+1)≤ρ3(a,1)=a21. We just need to prove the lower bound. For each hypergraph H∈Ha, let H⌊γp⌋+1 be the blow-up of H by replacing each vertex of P with 2a2+a+1a(⌊γp⌋+1) vertices and replacing each vertex of Q with 2a2+a+1⌊γp⌋+1 vertices. Let Ha⌊γp⌋+1={H⌊γp⌋+1:H∈Ha}. By Theorem 3.8, π(Ha⌊γp⌋+1)=π(Ha)=1−a21.
For large p, α(H⌊γp⌋+1)≤2a2+a+1(2a+1)(⌊γp⌋+1)+O(1)<p+1. By (1), t3(⌊γp⌋+1,p+1)≥1−π(Ha⌊γp⌋+1)=a21.
∎
Vaughan [22] computed an upper bound of π(K63) by flag algebra.
Theorem 3.19** (Vaughan, [22], [2]).**
π(K63)≤0.8583903.**
Combining Theorem 3.17, Theorem 3.18 and Theorem 3.19, we have the following conclusion.
Theorem 3.20**.**
Let γ be a real number with 3≤γ<722, there exist a postive integer p0=p(γ), such that for all integers p≥p0,
[TABLE]
Remark 3.21**.**
Combining Theorem 3.17 and Theorem 3.18, if π(K2a3)≤1−a21, then for any real number γ with a≤γ<a+2a+11, there exist a postive integer p0=p(γ), such that for all integers p≥p0, t3(⌊γp⌋+1,p+1)=a21. If Turán’s Conjecture (Conjecture 1.2) holds for odd number k=2a+1, that is π(K2a+13)=1−a21, then for any real number γ with a≤γ<a+21, there exist a postive integer p0=p(γ), such that for all integers p≥p0, t3(⌊γp⌋+1,p+1)=a21.
4 Concluding remarks
In this paper, we determine some exact values of local Turán densities and answer Question 1.6 partially; in particular, our results imply that the equality in Question 1.6 about exact values of the limits does not hold in general.
The local Turán density problems are still widely open. Here we discuss some natural problems.
Frankl, Huang and Rödl [11] extended Theorem 1.5 to r-uniform hypergraphs and made the following conjecture.
Conjecture 4.1** (Frankl-Huang-Rödl, [11]).**
For integers r≥2 and p sufficiently large,
[TABLE]
Note that Question 1.6 is a generalization of Theorem 1.4 and closely related to Conjecture 4.1.
Based on Theorem 1.7, we extend Question 1.6 to the following more general question. Recall that, for an integer r≥3 and a real number γ>1, we defined the family Fγr of r-uniform hypergraphs as
Fγr={r-uniform graphs F:ν(F)>γα(F)}.
Question 4.2**.**
Let r≥3 be an integer and γ>1 be a real number. Does there exist a positive integer p0=p0(r,γ), such that for all integers p≥p0,
[TABLE]
Using the blow-up theorem of Brown and Simonovits [4], one can show that Question 4.2 is equivalent to the following.
Question 4.3**.**
Let r≥3 be an integer and γ>1 be a real number. Does there exist a finite subfamily Ffin⊂Fγr such that
[TABLE]
We say a family H of r-uniform hypergraphs compact if there is a finite subfamily Hfin⊂H such that π(Hfin)=π(H) (See Conjecture 1 in [4]). In this notion, Question 4.3 asks whether the family Fγr is compact. The result of Frankl and Stechkin [13] shows that the family Fγr is compact when 1≤γ<r−1r. Theorem 1.5 shows that the family F23 is compact.
Moreover, the corresponding hypergraph family of each case considered in Theorem 1.8 is also compact.
Acknowledgment. Very recently we learn that Frankl and Nie [12] have also obtained results in this topic, which partially overlap with our Theorem 1.8 when r=3 and 2≤γ<7/3.